Repository navigation
Preserve distinct mutation output names in graph signatures (#24105) - #24105
Open
ThomasJannaud wants to merge 2 commits into
Open
ThomasJannaud wants to merge 2 commits into
ThomasJannaud wants to merge 2 commits into
Conversation
ThomasJannaud
requested review from
JacobSzwejbka and
larryliu0820
as code owners
October 23, 2026 05:07
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/24105
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 Unclassified FailureAs of commit 63a1c00 with merge base 185668b ( UNCLASSIFIED FAILURE - DrCI could not classify the following job because the workflow did not run on the merge base. The failure may be pre-existing on trunk or introduced by this PR:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
Contributor
|
@ThomasJannaud has exported this pull request. If you are a Meta employee, you can view the originating Diff in D127072149. |
This PR needs a
|
meta-codesync
Bot
changed the title
ThomasJannaud
added a commit
to ThomasJannaud/executorch
that referenced
this pull request
Oct 23, 2026
…24105) Summary: Problem `ExportGraphSignature.buffers_to_mutate` and `user_inputs_to_mutate` index writes by output node name. No-op folding and CSE can merge two writes into one node, so one of the mutation entries is lost from the signature. Solution Rebuild the signature from `output_specs` (`_get_updated_graph_signature` in `exir/_program_utils.py`). Copy a computed result only when two writes collide in the same mutation map, and keep positional user outputs and ordinary graph uses shared. Select ATen, edge or dim-order cloning from the program's verifier. Preserve value/debug metadata and FX provenance on the copies without inheriting allocation or custom-kernel placement annotations. Mutation output copies keep independent storage, as required by the existing output pointer contract. This is the ExecuTorch half of D125489522, split out so that ExecuTorch owners can review it on its own. The Helios half (removing the Helios-side duplicated-output splitter that this change makes unnecessary) stays in D125489522, stacked on top. fbcode and xplat copies are identical. {F2001970137} Differential Revision: D127072149
ThomasJannaud
force-pushed
the
export-D127072149
branch
2 times, most recently
from
October 23, 2026 15:28
aa2f8a4 to
8c4bc17
Compare
ThomasJannaud
added a commit
to ThomasJannaud/executorch
that referenced
this pull request
Oct 23, 2026
…24105) Summary: Problem `ExportGraphSignature.buffers_to_mutate` and `user_inputs_to_mutate` index writes by output node name. No-op folding and CSE can merge two writes into one node, so one of the mutation entries is lost from the signature. Solution Rebuild the signature from `output_specs` (`_get_updated_graph_signature` in `exir/_program_utils.py`). Copy a computed result only when two writes collide in the same mutation map, and keep positional user outputs and ordinary graph uses shared. Select ATen, edge or dim-order cloning from the program's verifier. Preserve value/debug metadata and FX provenance on the copies without inheriting allocation or custom-kernel placement annotations. Mutation output copies keep independent storage, as required by the existing output pointer contract. This is the ExecuTorch half of D125489522, split out so that ExecuTorch owners can review it on its own. The Helios half (removing the Helios-side duplicated-output splitter that this change makes unnecessary) stays in D125489522, stacked on top. fbcode and xplat copies are identical. {F2001970137} Differential Revision: D127072149
ThomasJannaud
added a commit
to ThomasJannaud/executorch
that referenced
this pull request
Oct 23, 2026
…24105) Summary: Problem `ExportGraphSignature.buffers_to_mutate` and `user_inputs_to_mutate` index writes by output node name. No-op folding and CSE can merge two writes into one node, so one of the mutation entries is lost from the signature. Solution Rebuild the signature from `output_specs` (`_get_updated_graph_signature` in `exir/_program_utils.py`). Copy a computed result only when two writes collide in the same mutation map, and keep positional user outputs and ordinary graph uses shared. Select ATen, edge or dim-order cloning from the program's verifier. Preserve value/debug metadata and FX provenance on the copies without inheriting allocation or custom-kernel placement annotations. Mutation output copies keep independent storage, as required by the existing output pointer contract. This is the ExecuTorch half of D125489522, split out so that ExecuTorch owners can review it on its own. The Helios half (removing the Helios-side duplicated-output splitter that this change makes unnecessary) stays in D125489522, stacked on top. fbcode and xplat copies are identical. {F2001970137} Differential Revision: D127072149
ThomasJannaud
force-pushed
the
export-D127072149
branch
2 times, most recently
from
October 23, 2026 21:52
206f7ce to
50b94d4
Compare
ThomasJannaud
added a commit
to ThomasJannaud/executorch
that referenced
this pull request
Oct 23, 2026
…24105) Summary: Problem `ExportGraphSignature.buffers_to_mutate` and `user_inputs_to_mutate` index writes by output node name. No-op folding and CSE can merge two writes into one node, so one of the mutation entries is lost from the signature. Solution Rebuild the signature from `output_specs` (`_get_updated_graph_signature` in `exir/_program_utils.py`). Copy a computed result only when two writes collide in the same mutation map, and keep positional user outputs and ordinary graph uses shared. Select ATen, edge or dim-order cloning from the program's verifier. Preserve value/debug metadata and FX provenance on the copies without inheriting allocation or custom-kernel placement annotations. Mutation output copies keep independent storage, as required by the existing output pointer contract. This is the ExecuTorch half of D125489522, split out so that ExecuTorch owners can review it on its own. The Helios half (removing the Helios-side duplicated-output splitter that this change makes unnecessary) stays in D125489522, stacked on top. fbcode and xplat copies are identical. {F2001970137} Differential Revision: D127072149
Gasoonjia
added
ciflow/trunk
ciflow/android
Summary: Problem External constants can be represented by an integer address rather than a host pointer. Treating both representations as a directly accessible pointer can dereference an invalid address or accept a truncated/overflowing extent. Solution Distinguish pointer-backed and address-backed `FreeableBuffer` storage in the tensor parser, validate requested ranges and resolve address-backed storage through the loader. Keep fbcode/xplat mirrors and dual-address loader regressions together. Dependencies and priority Generic loader correctness change; independent of the mask, embedding splitting and hybrid optimizations. Shared resident views consume it. P0 for address-backed loading correctness. History provenance Resolve local external tensor data through checked pointer or integer addresses. Preserve both mirrored parsers and keep pointer-backed, overflow, short-storage, and shared `FlatTensor` loader regressions together. Differential Revision: D125508895
…24105) Summary: Problem `ExportGraphSignature.buffers_to_mutate` and `user_inputs_to_mutate` index writes by output node name. No-op folding and CSE can merge two writes into one node, so one of the mutation entries is lost from the signature. Solution Rebuild the signature from `output_specs` (`_get_updated_graph_signature` in `exir/_program_utils.py`). Copy a computed result only when two writes collide in the same mutation map, and keep positional user outputs and ordinary graph uses shared. Select ATen, edge or dim-order cloning from the program's verifier. Preserve value/debug metadata and FX provenance on the copies without inheriting allocation or custom-kernel placement annotations. Mutation output copies keep independent storage, as required by the existing output pointer contract. This is the ExecuTorch half of D125489522, split out so that ExecuTorch owners can review it on its own. The Helios half (removing the Helios-side duplicated-output splitter that this change makes unnecessary) stays in D125489522, stacked on top. fbcode and xplat copies are identical. {F2001970137} Differential Revision: D127072149
ThomasJannaud
force-pushed
the
export-D127072149
branch
from
October 24, 2026 16:11
50b94d4 to
63a1c00
Compare
This branch was successfully deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary:
Problem
ExportGraphSignature.buffers_to_mutateanduser_inputs_to_mutateindex writes by output node name. No-op folding and CSE can merge two writes into one node, so one of the mutation entries is lost from the signature.Solution
Rebuild the signature from
output_specs(_get_updated_graph_signatureinexir/_program_utils.py). Copy a computed result only when two writes collide in the same mutation map, and keep positional user outputs and ordinary graph uses shared. Select ATen, edge or dim-order cloning from the program's verifier. Preserve value/debug metadata and FX provenance on the copies without inheriting allocation or custom-kernel placement annotations. Mutation output copies keep independent storage, as required by the existing output pointer contract.This is the ExecuTorch half of D125489522, split out so that ExecuTorch owners can review it on its own. The Helios half (removing the Helios-side duplicated-output splitter that this change makes unnecessary) stays in D125489522, stacked on top. fbcode and xplat copies are identical.
{F2001970137}
Differential Revision: D127072149